Skills2026-08-11 · source-backed

Branch live instead of replaying logs when you evaluate model routing.

Issue

2026-08-11

Confidence

source-backed

Sources

Redaction

redacted

Story

If you're deciding between Haiku and Opus per step, fork the actual trajectory and re-run with the alternate model in a rebuilt environment. Log-stitching replay leaves only 3% of states valid and mispredicted every success-relevant outcome in the Replay Gap study. Your offline router benchmark is currently measuring a world that doesn't exist.

Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage
Four stacked tools cut Claude Code bills over 90%
Linked by a graph relationship (Claude Code uses Opus); both cover Haiku, Opus; overlapping topics (instead, model).
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage
A $0.25 Model Beat a $15 Model. The Secret Was the Skill, Not the Brain.
Linked by a graph relationship (Opus built by Anthropic); both cover Haiku, Opus; overlapping topics (haiku, model).
Claude Code uses Opus / Shared entity: Opus / Same source domain / Shared topic / Earlier coverage
Tencent Leaderboard Splits Same Model's Score by 13.5 Points
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; reported by the same outlet (arxiv.org).
Claude Code uses Opus / Shared entities / Earlier coverage / Tension
GPT-5.6 Sol, Terra, and Luna landed in GitHub Copilot the same day OpenAI got public rollout clearance.
Linked by a graph relationship (Claude Code uses Opus); both cover Haiku, Opus; earlier Haiku coverage from 2026-07-10.
Claude Code uses Opus / Shared entities / Same source domain
Wrapping the agent loop in a runtime-guidance DSL lifted skill success 24.1%, tested on Claude Code and Codex directly.
Linked by a graph relationship (Claude Code uses Opus); both cover Haiku, Opus; reported by the same outlet (arxiv.org).
Cursor uses Opus / Shared entity: Haiku / Same source domain / Earlier coverage / Tension
66.5% of Malicious GitHub Issues Get Past Every Guardrail in Cursor, Claude Code, and Codex
Linked by a graph relationship (Cursor uses Opus); both cover Haiku; reported by the same outlet (arxiv.org).
Claude Code uses Opus / Shared entity: Opus / Shared topic / Earlier coverage / Tension
Grok Build Costs $2.49 a Task, Opus 4.8 Costs $11.80
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (benchmark, model).
Claude Code uses Opus / Shared entity: Opus / Shared topic / Earlier coverage
GLM-5.2 is the hedge, and the benchmarks say it's a real one
Linked by a graph relationship (Claude Code uses Opus); both cover Opus; overlapping topics (benchmark, between, model).

Source trail

Replay Gap study

Entities

Branch Haiku Opus Replay Gap

Provenance

Canonical issue: Ramsay Research Agent — August 11, 2026
AI generated: no
Story unit: 2026-08-11-branch-live-instead-of-replaying-logs-when-you-evaluate-model-routing
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Branch live instead of replaying logs when you evaluate model routing.

Story

Related stories

Four stacked tools cut Claude Code bills over 90%

A $0.25 Model Beat a $15 Model. The Secret Was the Skill, Not the Brain.

Tencent Leaderboard Splits Same Model's Score by 13.5 Points

GPT-5.6 Sol, Terra, and Luna landed in GitHub Copilot the same day OpenAI got public rollout clearance.

Wrapping the agent loop in a runtime-guidance DSL lifted skill success 24.1%, tested on Claude Code and Codex directly.

66.5% of Malicious GitHub Issues Get Past Every Guardrail in Cursor, Claude Code, and Codex

Grok Build Costs $2.49 a Task, Opus 4.8 Costs $11.80

GLM-5.2 is the hedge, and the benchmarks say it's a real one

Source trail

Entities

Provenance

Branch live instead of replaying logs when you evaluate model routing.

Story

Related stories

Four stacked tools cut Claude Code bills over 90%

A $0.25 Model Beat a $15 Model. The Secret Was the Skill, Not the Brain.

Tencent Leaderboard Splits Same Model's Score by 13.5 Points

GPT-5.6 Sol, Terra, and Luna landed in GitHub Copilot the same day OpenAI got public rollout clearance.

Wrapping the agent loop in a runtime-guidance DSL lifted skill success 24.1%, tested on Claude Code and Codex directly.

66.5% of Malicious GitHub Issues Get Past Every Guardrail in Cursor, Claude Code, and Codex

Grok Build Costs $2.49 a Task, Opus 4.8 Costs $11.80

GLM-5.2 is the hedge, and the benchmarks say it's a real one

Source trail

Entities

Provenance